* Remove hook `torch.nan_to_num(x)`

Observed erratic training behavior (green line) with the nan_to_num hook in classifier branch. I'm going to remove it from master.

* Update train.py